Author's personal copy - web.cs.ucdavis.edu

An iterative closest point approach for the registrationof volumetric human retina image data obtained by opticalcoherence tomography

Xin Wang1,2 & Zhen-Long Zhao1 & Arlie G. Capps3 &

Bernd Hamann3

Received: 7 May 2015 /Revised: 10 December 2015 /Accepted: 26 January 2016# Springer Science+Business Media New York 2016

Abstract This paper introduces an improved approach for the volume data registration ofhuman retina. Volume data registration refers to calculating out a near-optimal transformationbetween two volumes with overlapping region and stitching them together. Iterative closestpoint (ICP) algorithm is a registration method that deals with registration between points.Classical ICP is time consuming and often traps in local minimum when the overlappingregion is not big enough. Optical Coherence Tomography (OCT) volume data are severalseparate, partially overlapping tiles. To stitch them together is a technology in computer aideddiagnosis. In this paper, a new 3D registration algorithm based on improved ICP is presented.First, the Canny edge detector is applied to generate the point cloud set of OCT images. Afterthe detection step, an initial registration method based on the feature points of the point cloudis proposed to determine an initial transformation matrix by using singular value decomposi-tion (SVD) method. Then, an improved ICP method is presented to accomplish fine registra-tion. Corresponding point in the point cloud is weighted to reduce the iteration times of ICPalgorithm. Finally, M-estimation is used as the objective function to decrease the impact ofoutliers. This registration algorithm is used to process human retinal OCT volume pairs thatcontain an overlapping region of 75 × 500 × 375 voxels approximately. Then a comparativeexperiment is conducted on some public-available datasets. The experimental results show thatthe proposed method outperforms the classical method.

Multimed Tools ApplDOI 10.1007/s11042-016-3302-9

* Xin [email protected]

1 College of Computer Science and Technology, Jilin University, 130012 Changchun, China2 Key Laboratory of Symbolic Computation and Knowledge Engineer of Ministry of Education, Jilin

University, 130012 Changchun, China3 Institute for Data Analysis and Visualization (IDAV), Department of Computer Science, University of

California, Davis, Davis, CA 95616-8562, USA

Author's personal copy

Keywords Volume data registration . Optical coherence tomography. Retinal image . Iterativeclosest point . Point cloud

1 Introduction

Optical coherence tomography fundus images, which can provide high-resolution cross-sectional information of the human retina, are indispensable for clinical diagnosis, treatmentand surgical evaluation of diseases such as macular degeneration and glaucoma. OCT pro-cesses high resolution in vivo volumetric imaging while the acquisition time is relatively shortfor avoiding motion artifacts caused by involuntary eye movement. Hence, the scan range ofOCT is limited and only small volumetric data is acquired during one scan [23]. Therefore, itwould be reasonable to focus on creating an OCT volume data covering a large field of view(FOV) [1].

Literatures on two-dimensional medical image registration are extensive [2, 16, 20]. KratikaSharma and Ajay Goyal [21] made a survey of image registration techniques and providedoverall source of recent as well as classic research. They divided image registration steps intofour categories as spatial relations method, relaxation methods, pyramids and wavelets as wellas methods using invariant descriptors. Mei-sen Pan et al. [18] proposed an image registrationmethod based on edges that detected by the B-spline gradient operator. This method has afairly simple implementation and it can be adapted to both mono-modality and multi-modalityimage registrations. Moreover, invariant descriptors can also serve in two-dimensional regis-tration. Lucian Ciobanu and Luís Côrte-Real [7] provided a solution to register two complete-overlapped views problem based on iterative filtering of SIFT-generated key point matches,using the Hough transform and blocking matching. An iteration based approach was used toeliminate the most probable outlier and rebuilding the relations. It makes an overall significantreduction of the outliers while maintaining a high rate of correct matches. To achieve accurateand robust registration, another novel idea is modeling the entire image distribution. ShihuiYing et al. [24] first introduced this concept. Thus the procedure of groupwise registration isformulated as the dynamic shrinkage of graph on the manifold which brings the advantage ofpreserving the topology of the image distribution during the groupwise registration. Besides,some other feature based registration approaches can obtain satisfactory results [4, 15].

However, OCT volume data are composed of two dimensional B-scan images. Aboveregistration methods may encounter difficulties in calculating memory and time limitationduring process these volume. To register 3D OCT volume data, a novel idea to obtain a largeFOVof OCT images is to create a montage [14]. In this method, blood vessel ridges are usedas the feature of interest and a procedure based on resampling, interpolation, and cross-correlation is proposed to piece together the full OCT data. The montage method can integratethe dispersed, partially overlapping OCT images into a large 3D OCT image. However, thismethod would fail to register when blood vessel ridges are fuzzy. Other strategies generate awide-field volume by using existing tools and platforms. Meng Lu [17] proposed an acceler-ation method based on Compute Unified Device Architecture (CUDA) which created byNVIDIA. This algorithm can improve the performance of 3D medical image registration andaccelerate the calculation speed as well, which is suitable for large-scale data processing.Stephan Preibisch et al. [19] implement a stitching plugin in ImageJ that reconstructs severaltypes of tiled microscopy acquisitions ranging from mosaics of histological 2D images to setsof gray-scale and RGB 3D confocal stacks. No prior knowledge is required and brightness

Multimed Tools Appl


differences between tiles are compensated by a smooth intensity transition. In addition toabove these, some studies focus on stitching software and related stitching tools are developedsuccessively [5, 10, 25]. However, some subtle non-rigid transformation appears during scanprocedure due to the instability of ophthalmic instruments and involuntary eye movement. Inhence these methods, which mainly deal with rigid transformation, have limitations inprocessing clinical ophthalmology OCT images.

In this paper, we focus on proposing a registration platform that can process non-rigidtransformation and generate a large FOVof OCT volumetric data quickly and accurately. Weuse a coarse-to-fine strategy to calculate the transformation matrix which could integratevolumetric data together. First, edge points of each retinal image are selected by Canny edgedetector and these edge points are collected together as point cloud. The purpose of Cannyedge detector is to reduce the amount of points in point cloud and exclude the impact of noise.Then a method based on the feature points of point cloud is proposed to calculate the initialrough rigid registration matrix. Finally, the fine registration matrix is calculated by animproved ICP method. This method for global optimization could handle non-rigid transfor-mation in its iterating step. Weighting method is taken into account when calculating thedistance of each corresponding points. M-estimation object function is introduced to eliminatethe abnormal points.

Our improved method accomplished 3D retinal OCT volume data registration and success-fully broke through the efficiency bottleneck of volume registration. By comparison, timeconsumption and registration accuracy of our method are satisfactory.

The remainder of this paper is structured as follows. In the section 2, we give a detailsdescription of the proposed registration approach with the sub-sections 2.1, 2.2 and 2.3. Thesection 3 highlights the experiment implementation details and presents the results of theproposed approach on retinal OCT sub-volumes and some public-available datasets.Comparison with other registration method is made in the section 4 on a wide range ofOCT datasets. Finally, a review of this paper and future work are presented in the section 5.

2 Materials and method

For the purpose of this paper, we use two 3D image sets as Reference Set and Target Set.Figure 1 shows these two sets. They are adjacent sub-volumes of human retina acquired byOCT instrument. We use a fundus image to show the actual position of the two sub-volumes.

The aim is to find out proper transformation matrixes to integrate all these sets into a fullOCT volume data covering a large FOV. To begin with, the schematic of our algorithm issummarized as Fig. 2.

Canny method in first phase solved the problem of the large amount of data. In theinitial registration phase, SVD method is used to decompose the feature points andeliminate the translation and rotation misalignment. Two constraints are added to im-prove the time consumption and registration accuracy of classical ICP method in Fineregistration phase.

2.1 Generate point clouds of volumetric images

3D point clouds refer to a set of spatial data points and usually are used to represent theexternal surface of an object. Current commercial OCT instrument can obtain tiny volumetric

Multimed Tools Appl


images with high resolution. Generating point clouds by original OCT images will produce ahuge amount of point clouds, which beyond the processing ability of current hardware andalgorithms. Besides, during data acquisition procedure, there will generate approximately0.1 % ~ 5 % noise points. These noise points will affect registration process and result inaccuracy decline. Thus, Canny method is used to detect the edge of retinal images, which helpsto accelerate registration speed and reduce calculation amount. Canny method uses dualthreshold to gather new edges in an 8-adjacement area, by which single noise point will notbe treated as a part of edge. Besides, the edge detected by Canny method is the featureextraction of B-scans, which does not actually change the position information of the over-lapping region of OCT volume, so using Canny method will lead to a good registrationcompared to original datasets. Figure 3 shows the result of a B-scan retinal image processed byCanny method.

Each pixel of the edge is regarded as a spatial point and all the edges of Reference Set andTarget Set are adopted as the reference point cloud SR and target point cloud ST, respectively.

R T

Fig. 1 The actual position ofReference Set (left) and Target Set(right). Fundus image is used forreference. Note that the two setsare adjacent human retina andcomposed by the superposition ofsingle B-scan obtained by OCTinstrument

Multimed Tools Appl


Canny edge detection eliminates the impact of noise and decreases the size of point cloud,thereby reducing the calculation burden of the numerous OCT fundus images.

Volumetric images

Edge refine and extraction

Point clouds

Calculate feature points

Decompose by SVD

Rotationmatrix R

Translation matrix T

Initialvolume

Corresponding points weighting

Objectivefunction withM-estimation

Volume withlarge FOV

Generation of point clouds

Initialregistration

Fineregistration

Finalresult

Fig. 2 Schematic diagram of the proposed algorithm. A coarse-to-fine transformation strategy is utilized toobtain OCT volume with large FOV

Fig. 3 Edge refine and extraction by Canny method (threshold λ1 = 300, λ2 = 900). The left image is a single B-scan retinal image acquired by OCT instrument, while the right shows the Canny edge of retinal image

Multimed Tools Appl


2.2 Initial registration

We use a coarse-to-fine strategy to work out the transformation matrix, considering bothregistration results and efficiency. The purpose of initial registration is to eliminate thetranslation and rotation misalignment and provide a favorable initial state for fine registration.To this end, feature points are extracted from the two point cloud datasets and SVD method isutilized to work out the rotation matrix R and translation vector T.

2.2.1 Extraction of feature points

To extract feature points, we first divide the point cloud sets into several spatial grids. Then wedelineate all the boundary grids based on a novel selection algorithm. Finally, we split theboundary grids of point cloud and extract the feature points from these grids.

Oriented boundary box is used to find the minimum bounding box of point cloud datasets.Oriented boundary box is an oriented algorithm that makes the bounding box smallest.Different from axis-aligned bounding boxes, oriented boundary box is defined as a cuboidwhose direction is arbitrary. After calculating the minimum bounding boxes of the two pointcloud sets SR and ST, scaling transformation is used to ensure the two minimum boundingboxes a roughly equal size. Then, the size of spatial grids can be obtained from the minimumbounding box, which is defined as:

Sgrid ¼ K•VQ

ð1Þ

where Q indicates the quantity of point cloud, V represents the volume of minimum boundingbox. K is a variable parameter. For dense point cloud, experiments show that when K isassigned to be 8 ~ 24, the size of spatial grid contains adequate points while will not be toolarge to affect registration accuracy. Let the size of spatial grid Sgrid be K times of the reciprocalof the point cloud density and then divide the minimum bounding box equally spaced alongthe three axes according to this size. So far, all points in point cloud belong to certain gridsaccording to their space coordinate. We define these spatial grids as occupied grid and emptygrid according to whether contains points. Next, we extract the boundary grids using thefollowing equation:

U x; y; zð Þ ¼ f x−1; y; zð Þ• f xþ 1; y; zð Þþf x; y−1; zð Þ• f x; yþ 1; zð Þþf x; y; z−1ð Þ• f x; y; zþ 1ð Þ

ð2Þ

where (x,y,z) is the spatial coordinate of a grid and f(x,y,z) represents the type of a grid. If a gridis a occupied grid, f(x,y,z) = 1, otherwise f(x,y,z) = 0. A spatial grid (x,y,z) is classified asboundary grid or not by calculating its six adjacent neighbors. U(x,y,z) is the sum of threeproducts and each product value is 0 or 1. For example, if f(x + 1,y,z) · f(x-1,y,z) = 1, means thatthe left gird (x + 1,y,z) and the right grid (x-1,y,z) of the current grid are both occupied grids.From the following figure (Fig. 4), we can conclude that a spatial grid (x,y,z) is a boundary gridwhen U(x,y,z) ≤ 1,which represents there are no more than four adjacent neighbor grids areoccupied grids. All boundary grids are selected by this method and each point in these grids isextracted as feature points.

Multimed Tools Appl


2.2.2 Registration based on feature points

Once we have obtained the feature points, we gather them to participate in the initialregistration. In our method, singular value decomposition is used to work out the rotationmatrix R and translation vector T between corresponding pairs. A target matrix is defined by

E ¼ 1M

XM

i¼1

Ri−CRð Þ % Ti−CTð ÞT ð3Þ

where CR and CT is the centroid of SR and ST respectively,M refers to the minimum number ofpoints between SR and ST while Ri and Ti represents ith point in SR and ST. Decompose E bySVD, Eq. (3) reduces to E=UDVT, the columns of U are the eigenvectors of the EET matrixand the columns of V are the eigenvectors of the ETE matrix. VT is the transpose of V while Drepresents a diagonal matrix, by definition the non-diagonal elements are zero. D= diag(di)with di ≥ di+1. Let

P ¼ I3 det Uð Þdet Vð Þ≥0diag 1; 1;−1ð Þ det Uð Þdet Vð Þ < 0

!ð4Þ

If rank(P) ≥2, then rotation matrix R can be calculated as R=UPVT and translation vectorT can be calculated as T=CT-RCR. Apply R and T to SR, we have finished the initialregistration step.

2.3 Fine registration

Assume SR and ST contains NR and NT points respectively. Time complexity of classical ICPalgorithm is O(NR · NT) (O(NR · logNT) at best). When processing large volume data, massivetime is spent on calculating the Euclidean distance between corresponding pairs. Anotherreason that makes classical ICP method unsuitable for our experiment is that it assumes thenearest point as corresponding point, which may trap the algorithm in local minimum. In thispaper, an improved ICP method is proposed. First, all the corresponding points are weightedand those whose weight is smaller than a given threshold are eliminated. Second, M-estimationis introduced into the objective function to decrease the impact of abnormal points. In classicalICP algorithm, all the points are given an equal weight. In hence, each point in the point cloudwill participate in the calculation of distance, which is the bottleneck of efficiency. In our

Fig. 4 Three cases of a boundary grid: a Boundary grid in a plane. b Boundary grid at the edge. c Boundary gridof vertex

Multimed Tools Appl


method, a linked list is maintained which stores the effective points in distance calculation.Points are classified as effective points when their weight is larger than the threshold. Weassume PT is one point in ST, the weight of PT has the form:

weight ¼ DisMAX

Dis PR;PTð Þ −1 ð5Þ

where PR is the current calculating point in SR, Dis(PR,PT) represents the Euclidean distancebetween PR and PT after initial registration step while DisMAX refers to the maximum distancebetween corresponding pairs. Effective points are stored in a linked list which will update afteriterating a point. Points are excluded if their weights are smaller than a fixed threshold ε whichis a variable argument that trades off time consumption and registration accuracy. If thethreshold ε is decreased, the registration procedure would be more accurate but time consum-ing, which means there are more points should be processed as effective ones. And onlyeffective points which are treated as corresponding points would participate in the calculationof Euclidean distance.

After excluding the points that have little effect on distance calculating, we introduce M-estimation to improve the objective function. M-estimation was proposed by Huber [12],which mainly deals with abnormal points. This method has overcome the shortcomings oftraditional methods that may have no solution. M-estimation is the process of finding anestimate value X which makes the residual error smallest. Experiments show that it is likely tomake registration trap into local minima when the number of abnormal points increases. Inorder to improve the robustness of registration algorithm, in this paper, a selecting weightiteration method proposed by Huber is used to reduce the sensitivity of the abnormal points.Huber defined the weighting factor of a point as:

ωυ ¼1 υj j≤ccυ

υj j > c

(

ð6Þ

where ν means the residue of a point, c is a constant. In general, c = 2σ(σ is the standarddeviation of a point in our algorithm). Huber M-estimation is a classical least squaresestimation when ν range from –c to c. Nevertheless, if residue ν is greater than c, the weightingfactor decreases while residue increasing. Equation (7) shows the equation to calculate theEuclidean square distance of a corresponding pair A and B.

DAB ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiωX • XA−XBð Þ2 þ ωY • YA−YBð Þ2 þ ωZ• ZA−ZBð Þ2

qð7Þ

We assume A, B are a corresponding pair of point cloud sets. ωx, ωy, ωz represents theweighting factor of Huber M-estimation respectively while XA, YA, ZA, XB, YB, ZB refers to thespatial coordinates of A and B, respectively.

The process of our improved ICP algorithm is summarized as follow:Given two point cloud sets SR and ST, an accuracy threshold τ, iterating the following steps:

– Exclude the points that have low weight according to a linked list (Initialized as empty).– Calculate all Euclidean distance between the two sets from Eq. (7).– For each point in SR, find the nearest point in ST as corresponding pairs and group them

together as the nearest point set ST1.

Multimed Tools Appl


– Calculate the translation vector T and rotation matrix R between SR and ST1 using least

mean square algorithm.– Apply registration matrix R and T to SR and get a new point cloud set SR

1. Update thelinked list and the root-mean-square error according to the new sets SR

1 and ST1.

Until the root-mean-square error converges to the given threshold τ.

3 Experiments and results

In order to evaluate the registration performance of our proposed method, we conduct severalexperiments on both clinical datasets of human retinal OCT sub-volumes and public-availabledatasets. For the sake of comparison, we test our algorithm on different size of point cloud anddiscuss the experimental results.

Our proposed algorithm has been applied to process two retinal OCT sub-volumes, whichare adjacent parts of the human retina structure. There are approximately 75 × 500 × 375 pixelsoverlapping region between the two sets. Figure 5 shows the iterating process from fourdifferent angles, the left visible area shows the initial position of the two point cloud (red pointsrepresent reference point cloud SR and green points represent target point cloud ST) while theright area indicates the real-time registration results of certain iteration step.

In our experiment, only the overlapping region of Reference Set and Target Set wasselected to participate in the generation of point cloud considering the calculation time(177,489 cloud points after Canny method). The experimental result demonstrates a relativelyaccurate registration of OCT fundus volume data. Figure 6 shows the result of our methodabout the point cloud sets in detail, the left image demonstrates the relative position of the twoOCT image sets. An obvious misalignment and some subtle deformation can be observed atzoom-in part. The proposed improved ICP method successfully registers the improper spatialcloud points which are illustrated by the right image in Fig. 6.

Fig. 5 Iterating process from different angles. The two point cloud sets are sampled as the left and the rights arereal-time registration results. The cube profile on the right is the oriented boundary box of point cloud

Multimed Tools Appl


The visualized experimental results are finally rendered by ImageJ. Figure 7 visualizes theinitial experimental data sets as well as the experimental result. There are four rendered OCTvolumes, the first two are experimental data that represent the Reference Set and Target Setmentioned above while the last two volumes on behalf of experimental result. Traditionalregistration of OCT volumes like that used in this paper may cause layer abruption especiallyin inner nuclear layer, photoreceptor cell layer and retinal pigment epithelium. However, theresult images in Fig. 7 demonstrate a relative satisfactory retinal OCT volume. No obviousmosaic trace is found in our experimental result even at the overlapping region, thus theobtained volume with large FOV may better help clinicians in the prevention and diagnosis ofophthalmology disease.

As for the performance of registration method, there are no quantified standard withabsolute certainty yet [8]. Researchers have proposed identification approaches from old andclassical [11] to novel [9, 22]. In this paper, we utilize registration error to evaluate theperformance, which is defined as:

ξ ¼ 1−

XN

i¼1

Success PR;PTð Þ

N•100% ð8Þ

Success(PR,PT) has the form:

Success PR;PTð Þ ¼ 1 Dis PR;PTð Þ≤δ0 Dis PR;PTð Þ > δ

!ð9Þ

where N indicates the total number of corresponding pairs, (PR,PT) is a corresponding pair.Success(PR,PT) indicates the registration result of corresponding pair (PR,PT).Success(PR,PT)= 1 when the Euclidean distance of the corresponding pair is smaller thanthe threshold δ. In the proposed performance evaluation formula (8), assuming that δ is 0.15times of the original Euclidean distance of (PR,PT) before registration, which means allcorresponding pairs whose Euclidean distance is less than 15 % of the original distance aresuccessfully registered. Table 1 shows the time consumption and registration error of ouralgorithm and classical ICP algorithm proposed by Besl [3].

Fig. 6 Partially results of overlapping point cloud. The first image shows a zoom part of the overlapping regionbefore registration, while the second shows the result of this part after using our method

Multimed Tools Appl


4 Discussion

With the development of OCT technique, 3D OCT volume data occupies an important place incomputer-aided diagnose. Therefore, a satisfactory high-resolution OCT volume data withlarge FOV would be clinical desired. In this area, our team reached some results. Dae Yu Kimet al. [13] reported high-speed acquisition at 125 kHz A-scans with phase-variance OCT thatcould reduce motion artifacts and increase the scanning area. Arlie G. Capps et al. [6]described a method for combining and visualizing a set of overlapping volume images withhigh resolution but limited spatial extent. Robert J. Zawadzki et al. [26] presented a shortreview of adaptive optics OCT instruments and proposed a method for correcting motionartifacts in adaptive optics OCT volume data.

In this paper, we propose an algorithm to integrate 3D OCT datasets. There are some newfeatures in our algorithm. Canny edge detector is applied to each OCT fundus image, whichcan remove the noise impact and reduce the calculation burden effectively (14,062,500 pointsof original overlapping region, 75 images, 375 × 500 pixels. 177,489 points after Canny edgedetection method). In the initial registration step, spatial grid partition and singular valuedecomposition method is used to find out the matrix of rigid transformation which may causeby involuntary eye movement during data acquisition. The core of our method is the improvedICP algorithm. Points are weighted by their distance to the current corresponding point. Thus,the points that have low weight will not participate in the iteration step. Besides, M-estimation

Table 1 Comparison of experimental result

Registration method Time consumption/s Registration error

Classical ICP 308.881 0.00096834

Initial registration 8.623 –

Improved ICP 82.517 0.00020539

Illustration of stitching performance on tiled volumetric images computed on a Windows machine with Intel® 2-Core CPU (2.93 GHz). Single tile dimension is 75 × 500 × 375

Fig. 7 Experimental data and results rendered in ImageJ. After iterating the points of overlapping region, we gotthe transformation matrixes and applied them to Reference Set and Target Set. a The Reference Set ofexperimental data. b The Target Set of experimental data. c The side view of the result. This image shows theretinal layers situation after registration. d The top view of the result. This image shows the situation of foveacentral is after registration

Multimed Tools Appl


is added to the equation that calculates the Euclidean square distance of corresponding pairs.There are three control parameters on x-axis, y-axis and z-axis respectively, which eliminatethe impact of outliers and make the algorithm robust. Comparing with classical ICP algorithm,these new constraints made our method a less time consumption and better accuracy (Table 1).Our approach consumed 91.140 s totally in the experiment, which shortened a lot thanclassical ICP; the registration accuracy has considerable improvement as well. In the initialregistration step, we eliminate the translation and rotation misalignment, which resulted in alarger overlapping region and fewer registration mistakes of corresponding points in theiterative process of ICP. Besides, the use of M-estimation also makes our approach moreaccurate. Weighting method is the core that makes our approach less time-consuming, whichreduces the amount of points in the calculation of distance.

We also test our algorithm on different sets of point cloud including some public-availabledatasets. For example, the first point cloud in Table 2 is the BStanford Bunny^ after beingprocessed by our method while the second one is the BDragon^ from The Stanford 3DScanning Repository (http://graphics.stanford.edu/data/3Dscanrep/). Table 2 and Fig. 8shows the performance of our improved ICP algorithm and classical ICP algorithm.

As is shown in Table 2 and Fig. 8, our method has obvious advantages when dealingwith large volume data. In most cases, our algorithm increases the efficiency of classical

Table 2 Different size of point cloud registration

Point cloud size Time consumption Time consumption

Classical ICP /s Improved ICP /s

9731 11.023 3.994

28107 42.928 19.562

69670 150.447 41.090

177489 308.881 82.517

385316 512.848 123.962

0 0.5 1 1.5 2 2.5 3 3.5 4

x 105

0

100

200

300

400

500

600

Point cloud size

Tim

e co

nsum

ptio

n

Classical ICP

Improved ICP

Fig. 8 Comparison of improved ICP with Besl’s method in time consumption with different point cloud size.With data size growing, our improved method stands out

Multimed Tools Appl


ICP algorithm by 70 %. Above analysis demonstrates the effectiveness of our methodand the possibility to apply it to process other OCT datasets that can be transformed intopoint clouds.

5 Conclusion

We present a non-rigid registration method of OCT retinal images, which can generate a 3Docular fundus volume covering a large FOV. Canny edge detection method has been applied inthe first stage of the proposed method to generate the point cloud. Oriented boundary box isused to process the point cloud set and obtain the feature points. The initial registration matrixis calculated based on these feature points and SVD method. At last, an improved ICPalgorithm is proposed to work out the fine registration matrix between the point cloud sets.Several human retinal OCT image sets and some public-available datasets are used to test theperformance of the proposed method. The experimental results show the performance of ouralgorithm has an obvious improvement compared with classical ICP algorithm in terms of timeconsumption and registration accuracy. In clinical practice, there are several partially overlap-ping OCT volumes of human retina. Under the same coordinate system, each volume has afixed coordinate. So after registering of two volumes, we regard them as a new volume andintegrate it with other volumes. The proposed method could provide strong support for clinicaltreatment and diagnosis. Our future work will focus on a self-adapted strategy to register OCTsub-volumes automatically.

Acknowledgments This work was supported by the National Natural Science Foundation of China (No.60905022) and the Jilin Provincial Research Foundation for Basic Research, China (No. 201105016). We thankthe members of the Institute for Data Analysis and Visualization (IDAV) at the University of California, Davis.We also thank Jack Werner and Robert Zawadzki of the Vision Science and Advanced Retinal ImagingLaboratory at the University of California, Davis.

References

1. Assayag O, Antoine M, Sigal-Zafrani B et al (2014) Large field, high resolution full field optical coherencetomography: a pre-clinical study of human breast tissue and cancer assessment. Technol Cancer Res Treat13:455–468

2. Ayyachamy S, Manivannan VS (2013) Medical image registration based retrieval using distance metrics. IntJ Imaging Syst Technol 23:360–371

3. Besl PJ, McKay ND (1992) Method for registration of 3-D shapes. Robotics-DL tentative. Int Soc OptPhoton 586–606

4. Biswas B, Dey K N, Chakrabarti A (2015) Medical image registration based on grid matching usingHausdorff Distance and Near set. 2015 I.E. Eighth International Conference on Advances in PatternRecognition (ICAPR)1–5

5. Bria A, Silvestri L, Sacconi L et al (2012) Stitching terabyte-sized 3D images acquired in ConfocalUltramicroscopy. 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI): 1659–1662.doi:10.1109/ISBI.2012.6235896

6. Capps AG, Zawadzki RJ, Werner JS et al (2013) Combined volume registration and visualization. Vis MedLife Sci, Proc 7–11. doi: 10.2312/PE.VMLS.VMLS2013.007-011

7. Ciobanu L, Côrte-Real L (2011) Iterative filtering of SIFT keypoint matches for multi-view registration inDistributed Video Coding. Multimed Tools Appl 55:557–578

8. Cohen EAK, Ober RJ (2013) Analysis of point based image registration errors with applications in singlemolecule microscopy. IEEE Trans Signal Process 6291–6306

Multimed Tools Appl


9. Datteri RD, Liu Y, D’Haese P et al (2014) Validation of a non-rigid registration error detection algorithmusing clinical MRI brain data. IEEE Trans Med Imaging 34:86–96

10. Emmenlauer M, Ronneberger O, Ponti A et al (2009) XuvTools: free, fast and reliable stitching of large 3Ddatasets. J Microsc 233:42–60

11. Hemler PF, Napel S, Sumanaweera TS et al (1995) Registration error quantification of a surface-basedmultimodality image fusion system. Med Phys 22:1049–1056

12. Huber PJ (2009) Robust statistics, 2nd edn. Wiley, Hoboken13. Kim DY (2011) In vivo volumetric imaging of human retinal circulation with phase-variance optical

coherence tomography. Biomed Opt Express 2:1504–C151314. Li Y, Gregori G, Lam BL et al (2011) Automatic montage of SD-OCT data sets. Opt Express 19:26239–

2624815. Li Y, Stevenson R (2014) Incorporating global information in feature-based multimodal image registration. J

Electron Imaging 23(2):76–85. doi:10.1117/1.JEI.23.2.02301316. Liu B, Zhang B, Wan C et al (2014) A non-rigid registration method for cerebral DSA images based on

forward and inverse stretching–avoiding bilinear interpolation. Bio-Med Mater Eng 24:1149–115517. Meng L (2014) Acceleration method of 3D medical images registration based on compute unified device

architecture. Bio-Med Mater Eng 24:1109–111618. Pan M, Jiang J, Rong Q et al (2014) A modified medical image registration. Multimed Tools Appl 70:1585–

161519. Preibisch S, Saalfeld S, Tomancak P (2009) Globally optimal stitching of tiled 3D microscopic image

acquisitions. Bioinformatics 25:1463–146520. Riffi J, Mahraz AM, Tairi H (2013) Medical image registration based on fast and adaptive bidimensional

empirical mode decomposition. IET Image Process 7:567–57421. Sharma K, Goyal A (2013) Classification based survey of image registration methods. Int Conf Comput

Commun Netw Technol 1–722. Surucu M, Roeske J (2013) A novel metric to evaluate dose deformation error for deformable image

registration algorithms. Med Phys. doi:10.1118/1.481432223. Vignali L, Solinas E, Emanuele E (2014) Research and clinical applications of optical coherence tomography

in invasive cardiology: a review. Curr Cardiol Rev 10:369–37624. Ying S, Wu G, Wang Q, et al (2013) Groupwise registration via graph shrinkage on the image manifold.

IEEE Conf Comput Vis Pattern Recogn 2323–233025. Yu Y, Peng H (2011) Automated high speed stitching of large 3D microscopic images. Biomedical Imaging:

From Nano to Macro. 2011 I.E. International Symposium on IEEE:238–24126. Zawadzki RJ, Capps AG, Kim DY et al (2014) Progress on developing adaptive optics–optical coherence

tomography for in vivo retinal imaging: monitoring and correction of eye motion artifacts. IEEE J Sel TopQuantum Electron 20:7100912

Xin Wang master’s supervisor, Female, born in 1975, received the PhD degree in computer science and technologyfrom Jilin University in 2006. She worked in Jilin University from 1999. She also worked in University of California,Davis from 2011 to 2012 as a visiting scholar. Her research interests include image processing, computer graphics,bioinformatics and computational biology.

Multimed Tools Appl


Zhen-Long ZhaoMale, born in 1990, received the B.S. degree in software college from Jilin University, Changchun,China, in 2013. He is currently pursuing the master degree in computer science and technology at Jilin University. Hisresearch interests include image processing, computer graphics and bioinformatics.

Arlie G. CappsMale, received the B.S. degree in computer science from Brigham Young University, Provo, UT, USA,in 2004. He is currently pursuing the Ph.D. degree in computer science at the University of California, Davis, CA, USA.He is a Livermore Graduate Scholar at Lawrence Livermore National Laboratory. His research interests include scientificand medical volume visualization, multimodal data fusion, and error quantification and correction.

Bernd HamannMale, teaches computer science at the University of California, Davis. His main areas of interest aredata visualization, geometric design and computing, and computer graphics. He studied mathematics and computerscience at the Technical University of Braunschweig, Germany, and at Arizona State University, Tempe, U.S.A.

Multimed Tools Appl


Author's personal copy - web.cs.ucdavis.edu

Documents

Transcript of Author's personal copy - web.cs.ucdavis.edu